Enterprise shift to agentic AI
OpenAI says enterprise now accounts for over 40% of its revenue, and companies are buying agentic workflows—systems of specialised agents that execute parts of a process—rather than just model access. (decrypt.co) That suggests firms are paying for durable workflow orchestration—context, tool calls, audit trails—rather than raw chat capabilities. (digitaldan.me)
OpenAI said on April 8 that business customers now bring in more than 40% of its revenue, and it expects enterprise sales to catch up with consumer sales by the end of 2026. The shift is notable because ChatGPT started as a mass-market product, but the faster-growing money now appears to be inside companies. (openai.com) The product companies are buying is changing too. OpenAI said GPT-5.4 is seeing “record engagement” in agentic workflows, which means customers are paying for systems that do jobs across several steps instead of paying only for a chatbot that answers one prompt at a time. (openai.com) An agentic workflow is closer to a back office team than a search box. One model can read an email, a second can pull data from a file or database, and a third can draft the reply, while the software keeps the whole chain connected. (developers.openai.com) OpenAI’s own developer tools are built around that idea. Its Agents software development kit says agents can use tools, hand work to specialized agents, stream partial results, and keep a full trace of what happened during the run. (developers.openai.com) That “full trace” is a big part of why companies pay for this. A consumer can tolerate a clever answer with no paper trail, but a bank, insurer, or drug company usually needs logs, permissions, and a record of which tool touched which data. (openai.com) OpenAI’s Responses application programming interface pushes in the same direction. The official documentation says it supports stateful interactions, built-in web search, file search, computer use, and function calling, which turns the model from a text generator into a worker that can act inside software. (developers.openai.com) In February, OpenAI launched Frontier, a platform for building and managing these agents inside companies. OpenAI described Frontier as a system with shared context, onboarding, permissions, and governance, which are the boring pieces that matter when software moves from demo to payroll system or claims desk. (openai.com) The customer names now attached to that push are large, regulated firms. OpenAI said recent enterprise demand includes Goldman Sachs, Philips, and State Farm, and State Farm separately said on February 5 that it was joining the Frontier platform to improve customer experience and employee productivity. (openai.com) (newsroom.statefarm.com) The scale underneath this is also moving fast. OpenAI said last week that its application programming interfaces process more than 15 billion tokens per minute, which helps explain why the company is talking less about single chat sessions and more about always-on systems embedded in daily operations. (openai.com) So the sale is no longer just “rent our model.” The sale is “let our software coordinate context, tools, handoffs, and audit trails across a workflow,” which is a much stickier business than selling raw access to a smart text box. (openai.com) (developers.openai.com)